[freetype2] master 0f644f3 23/30: [sdf] Add functions to compute pixel edge distances.

[Werner LEMBERG]

branch: master
commit 0f644f38e939f3fa82a2a91e94d53c6e05750e44
Author: Anuj Verma <anujv@iitbhilai.ac.in>
Commit: Werner Lemberg <wl@gnu.org>

    [sdf] Add functions to compute pixel edge distances.
    
    * src/sdf/ftbsdf.c (compute_edge_distance, bsdf_approximate_edge):
    New functions.
---
 ChangeLog        |   7 ++
 src/sdf/ftbsdf.c | 256 ++++++++++++++++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 262 insertions(+), 1 deletion(-)

diff --git a/ChangeLog b/ChangeLog
index 86a6308..950d7d5 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,12 @@
 2020-08-20  Anuj Verma  <anujv@iitbhilai.ac.in>
 
+       [sdf] Add functions to compute pixel edge distances.
+
+       * src/sdf/ftbsdf.c (compute_edge_distance, bsdf_approximate_edge):
+       New functions.
+
+2020-08-20  Anuj Verma  <anujv@iitbhilai.ac.in>
+
        [sdf] Add function to find edge pixels in a grid of alpha values.
 
        * src/sdf/ftbsdf.c (bsdf_is_edge): New function.
diff --git a/src/sdf/ftbsdf.c b/src/sdf/ftbsdf.c
index 2229354..9dc6953 100644
--- a/src/sdf/ftbsdf.c
+++ b/src/sdf/ftbsdf.c
@@ -192,7 +192,7 @@
                 FT_Int  w,    /* width                     */
                 FT_Int  r )   /* rows                      */
   {
-    FT_Bool  is_edge       = 0; 
+    FT_Bool  is_edge       = 0;
     ED*      to_check      = NULL;
     FT_Int   num_neighbors = 0;
 
@@ -240,4 +240,258 @@
 #undef CHECK_NEIGHBOR
 
 
+  /**************************************************************************
+   *
+   * @Function:
+   *   compute_edge_distance
+   *
+   * @Description:
+   *   Approximate the outline and compute the distance from `current`
+   *   to the approximated outline.
+   *
+   * @Input:
+   *   current ::
+   *     Array of Euclidean distances.  `current` must point to the position
+   *     for which the distance is to be caculated.  We treat this array as
+   *     a two-dimensional array mapped to a one-dimensional array.
+   *
+   *   x ::
+   *     The x coordinate of the `current` parameter in the array.
+   *
+   *   y ::
+   *     The y coordinate of the `current` parameter in the array.
+   *
+   *   w ::
+   *     The width of the distances array.
+   *
+   *   r ::
+   *     Number of rows in the distances array.
+   *
+   * @Return:
+   *   A vector pointing to the approximate edge distance.
+   *
+   * @Note:
+   *   This is a computationally expensive function.  Try to reduce the
+   *   number of calls to this function.  Moreover, this must only be used
+   *   for edge pixel positions.
+   *
+   */
+  static FT_16D16_Vec
+  compute_edge_distance( ED*     current,
+                         FT_Int  x,
+                         FT_Int  y,
+                         FT_Int  w,
+                         FT_Int  r )
+  {
+    /*
+     * This function, based on the paper presented by Stefan Gustavson and
+     * Robin Strand, gets used to approximate edge distances from
+     * anti-aliased bitmaps.
+     *
+     * The algorithm is as follows.
+     *
+     * (1) In anti-aliased images, the pixel's alpha value is the coverage
+     *     of the pixel by the outline.  For example, if the alpha value is
+     *     0.5f we can assume that the outline passes through the center of
+     *     the pixel.
+     *
+     * (2) For this reason we can use that alpha value to approximate the real
+     *     distance of the pixel to edge pretty accurately.  A simple
+     *     approximation is `(0.5f - alpha)`, assuming that the outline is
+     *     parallel to the x or y~axis.  However, in this algorithm we use a
+     *     different approximation which is quite accurate even for
+     *     non-axis-aligned edges.
+     *
+     * (3) The only remaining piece of information that we cannot
+     *     approximate directly from the alpha is the direction of the edge. 
+     *     This is where we use Sobel's operator to compute the gradient of
+     *     the pixel.  The gradient give us a pretty good approximation of
+     *     the edge direction.  We use a 3x3 kernel filter to compute the
+     *     gradient.
+     *
+     * (4) After the above two steps we have both the direction and the
+     *     distance to the edge which is used to generate the Signed
+     *     Distance Field.
+     *
+     * References:
+     *
+     * - Anti-Aliased Euclidean Distance Transform:
+     *     http://weber.itn.liu.se/~stegu/aadist/edtaa_preprint.pdf
+     * - Sobel Operator:
+     *     https://en.wikipedia.org/wiki/Sobel_operator
+     */
+
+    FT_16D16_Vec  g = { 0, 0 };
+    FT_16D16      dist, current_alpha;
+    FT_16D16      a1, temp;
+    FT_16D16      gx, gy;
+    FT_16D16      alphas[9];
+
+
+    /* Since our spread cannot be 0, this condition */
+    /* can never be true.                           */
+    if ( x <= 0 || x >= w - 1 ||
+         y <= 0 || y >= r - 1 )
+      return g;
+
+    /* initialize the alphas */
+    alphas[0] = 256 * (FT_16D16)current[-w - 1].alpha;
+    alphas[1] = 256 * (FT_16D16)current[-w    ].alpha;
+    alphas[2] = 256 * (FT_16D16)current[-w + 1].alpha;
+    alphas[3] = 256 * (FT_16D16)current[    -1].alpha;
+    alphas[4] = 256 * (FT_16D16)current[     0].alpha;
+    alphas[5] = 256 * (FT_16D16)current[     1].alpha;
+    alphas[6] = 256 * (FT_16D16)current[ w - 1].alpha;
+    alphas[7] = 256 * (FT_16D16)current[ w    ].alpha;
+    alphas[8] = 256 * (FT_16D16)current[ w + 1].alpha;
+
+    current_alpha = alphas[4];
+
+    /* Compute the gradient using the Sobel operator. */
+    /* In this case we use the following 3x3 filters: */
+    /*                                                */
+    /* For x: |   -1     0   -1    |                  */
+    /*        | -root(2) 0 root(2) |                  */
+    /*        |    -1    0    1    |                  */
+    /*                                                */
+    /* For y: |   -1 -root(2) -1   |                  */
+    /*        |    0    0      0   |                  */
+    /*        |    1  root(2)  1   |                  */
+    /*                                                */
+    /* [Note]: 92681 is root(2) in 16.16 format.      */
+    g.x = -alphas[0] -
+           FT_MulFix( alphas[3], 92681 ) -
+           alphas[6] +
+           alphas[2] +
+           FT_MulFix( alphas[5], 92681 ) +
+           alphas[8];
+
+    g.y = -alphas[0] -
+           FT_MulFix( alphas[1], 92681 ) -
+           alphas[2] +
+           alphas[6] +
+           FT_MulFix( alphas[7], 92681 ) +
+           alphas[8];
+
+    FT_Vector_NormLen( &g );
+
+    /* The gradient gives us the direction of the    */
+    /* edge for the current pixel.  Once we have the */
+    /* approximate direction of the edge, we can     */
+    /* approximate the edge distance much better.    */
+
+    if ( g.x == 0 || g.y == 0 )
+      dist = ONE / 2 - alphas[4];
+    else
+    {
+      gx = g.x;
+      gy = g.y;
+
+      gx = FT_ABS( gx );
+      gy = FT_ABS( gy );
+
+      if ( gx < gy )
+      {
+        temp = gx;
+        gx   = gy;
+        gy   = temp;
+      }
+
+      a1 = FT_DivFix( gy, gx ) / 2;
+
+      if ( current_alpha < a1 )
+        dist = ( gx + gy ) / 2 -
+               square_root( 2 * FT_MulFix( gx,
+                                           FT_MulFix( gy,
+                                                      current_alpha ) ) );
+
+      else if ( current_alpha < ( ONE - a1 ) )
+        dist = FT_MulFix( ONE / 2 - current_alpha, gx );
+
+      else
+        dist = -( gx + gy ) / 2 +
+               square_root( 2 * FT_MulFix( gx,
+                                           FT_MulFix( gy,
+                                                      ONE - current_alpha ) ) 
);
+    }
+
+    g.x = FT_MulFix( g.x, dist );
+    g.y = FT_MulFix( g.y, dist );
+
+    return g;
+  }
+
+
+  /**************************************************************************
+   *
+   * @Function:
+   *   bsdf_approximate_edge
+   *
+   * @Description:
+   *   Loops over all the pixels and call `compute_edge_distance` only for
+   *   edge pixels.  This maked the process a lot faster since
+   *   `compute_edge_distance` uses functions such as `FT_Vector_NormLen',
+   *   which are quite slow.
+   *
+   * @InOut:
+   *   worker ::
+   *     Contains the distance map as well as all the relevant parameters
+   *     required by the function.
+   *
+   * @Return:
+   *   FreeType error, 0 means success.
+   *
+   * @Note:
+   *   The function directly manipulates `worker->distance_map`.
+   *
+   */
+  static FT_Error
+  bsdf_approximate_edge( BSDF_Worker*  worker )
+  {
+    FT_Error  error = FT_Err_Ok;
+    FT_Int    i, j;
+    FT_Int    index;
+    ED*       ed;
+
+
+    if ( !worker || !worker->distance_map )
+    {
+      error = FT_THROW( Invalid_Argument );
+      goto Exit;
+    }
+
+    ed = worker->distance_map;
+
+    for ( j = 0; j < worker->rows; j++ )
+    {
+      for ( i = 0; i < worker->width; i++ )
+      {
+        index = j * worker->width + i;
+
+        if ( bsdf_is_edge( worker->distance_map + index,
+                           i, j,
+                           worker->width,
+                           worker->rows ) )
+        {
+          /* approximate the edge distance for edge pixels */
+          ed[index].near = compute_edge_distance( ed + index,
+                                                  i, j,
+                                                  worker->width,
+                                                  worker->rows );
+          ed[index].dist = VECTOR_LENGTH_16D16( ed[index].near );
+        }
+        else
+        {
+          /* for non-edge pixels assign far away distances */
+          ed[index].dist   = 400 * ONE;
+          ed[index].near.x = 200 * ONE;
+          ed[index].near.y = 200 * ONE;
+        }
+      }
+    }
+
+  Exit:
+    return error;
+  }
+
 /* END */